license: other license_name: qwen-research license_link: https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE base_model: Qwen/Qwen-Image-2.1 pipeline_tag: text-to-image tags:
- qwen-image-2.1
- lora
- comfyui
- safetensors
- image-generation
Uncensored Omni LoRA v0.1 for Qwen Image 2.1 — ComfyUI fused
Improved using Qwen.
This repository contains a LoRA adapter, not a standalone diffusion model. It is the ComfyUI fused-FFN conversion of qwen21_full1024_r128-step00032000.safetensors (rank 128, alpha 128, step 32,000). The original adapter has separate img_mlp.gate_layer and img_mlp.proj updates. This release combines them into img_mlp.gate_up while preserving the effective LoRA update. Attention and img_mlp.out adapters are unchanged.
| File | Size | SHA256 |
|---|---|---|
qwen-image-21-uncensored-omni-lora-v01-comfy-fused.safetensors | 872,491,248 bytes | 3bb604a1f73f8a51fd1f5e88ec541c935dccb9adee673f69c2cc1fc0bc4d6924 |
Source LoRA SHA256: daabb55d0f64f886f7e62df14be164ac3c4418ba07caaaae8d2126db5157afdf.
ComfyUI setup
Tested with ComfyUI 0.39.0 and Comfy-Org's non-Turbo qwen_image_2.1_int8_convrot.safetensors. Put this repository's LoRA in ComfyUI/models/loras/, then select it in LoraLoaderModelOnly after UNETLoader. A starting strength is 0.8. The example uses Comfy-Org's qwen3vl_8b_bf16.safetensors text encoder and qwen_image_2.1_vae_bf16.safetensors VAE.
The tested configuration loaded 192 LoRA patches and generated all six example images with zero lora key not loaded warnings. See the API workflow and example details. The base diffusion model, text encoder, and VAE are not included here.
Check the diffusion model's exact filename. Some Qwen Image 2.1 files, including qwen-image-2.1-UC-int8_convrot.safetensors from abenzerps, store the FFN as separate gate_layer and proj layers. This fused LoRA was validated with the Comfy-Org INT8 file above; it is not a universal replacement for the original, unfused LoRA. See the ComfyUI FFN layout discussion.
Examples
All examples use the same prompt and seed for the base and LoRA images. Settings: 1024 × 1024, seed 42, 25 steps, CFG 4.0, er_sde sampler, simple scheduler, LoRA strength 0.8. They are ordinary adult portrait prompts intended to show that the adapter loads and changes the output; they do not evaluate every intended use case.
| Scene | Base INT8 | Base INT8 + fused LoRA 0.8 |
|---|---|---|
| Beach fashion | Base beach example | LoRA beach example |
| Black dress portrait | Base portrait example | LoRA portrait example |
| City street portrait | Base street example | LoRA street example |
License
This adapter is derived from Qwen/Qwen-Image-2.1 and is subject to the Qwen Research License. See the NOTICE for attribution and modification details. The upstream license limits use to non-commercial research; review its full terms before use or redistribution.